Bio-inspired Object Recognition via Saccadic Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object recognition algorithms face challenges in accurately identifying objects, especially when dealing with partial or distorted images, requiring high resolution, and needing large sample sets for comparison, which complicates image searching and is inefficient for real-time applications.
Innovation Solution
A bio-inspired model that emulates saccadic eye movements and extracts features using Principal Component Analysis (PCA) and feature extraction algorithms, processing images through a vertebrate-inspired pathway to identify objects based on vertebrate, fovea, and lateral geniculate nucleus features, allowing for recognition of whole objects from partial images and handling variations in scale and orientation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If geometric algorithms are used for object recognition, then object identification can be achieved, but the system requires high resolution images and large sample sets, increasing device complexity and processing time
Solution Approach 1:
The patent segments the object recognition process into multiple hierarchical levels: initial feature extraction from partial images, intermediate feature refinement, and final object identification. This multi-stage segmentation allows the system to achieve accurate recognition without requiring complete high-resolution images or large sample sets at each stage, thereby reducing overall system complexity while maintaining measurement precision.
Solution Approach 2:
The patent introduces a temporal dimension to the recognition process by utilizing video sequences and temporal information from multiple frames. This allows the system to accumulate information over time, achieving accurate object identification from partial or low-resolution inputs by aggregating data across multiple time points, thus avoiding the need for complex spatial processing of single high-resolution images.
2Measurement precision
If geometric algorithms require complete pictures for feature analysis, then accurate relative position and shape detection is possible, but the system cannot handle partial object inputs, reducing adaptability
Solution Approach 1:
The patent performs preliminary feature extraction and candidate generation from partial images before complete object information is available. By pre-processing and extracting salient features from incomplete inputs, the system creates initial hypotheses that are then refined as more information becomes available, enabling accurate feature detection even when starting with partial object views.
Solution Approach 2:
The patent implements a dynamic recognition process that adapts to the completeness of input data. The system adjusts its processing strategy based on whether partial or complete images are provided, dynamically selecting appropriate feature extraction methods and confidence thresholds. This dynamic approach maintains measurement precision across varying input conditions while maximizing adaptability to different image completeness levels.
3Adaptability or versatility
If image translation and scaling are performed to match sample dimensions, then geometric algorithms can process images, but processing time increases and real-time performance deteriorates
Solution Approach 1:
The patent changes the fundamental parameters of the recognition approach by working with feature representations rather than raw pixel data. Instead of translating and scaling entire images to match sample dimensions, the system extracts scale-invariant and rotation-invariant features that can be directly compared across different image geometries. This parameter transformation from spatial coordinates to feature space eliminates the need for computationally intensive image warping while maintaining adaptability to various image formats and achieving real-time processing speeds.
Data Source
AI summary
Methods, systems, and storage media are disclosed herein for determining objects in image data that match a plurality of captured images, in which one or more of the captured images from among the plurality of captured images may depict only a portion of an object. A computing device may extract a first feature from at least one block of the plurality of blocks from a captured image; determine a second feature; and determine a third feature. The computing device may determine, as a matching image, a stored image from among a plurality of stored images including an object that has greatest maximum correlation with the third feature. The computing device may similarly determine a matching image for each of the captured images of the plurality of images; and determine a final matching image from among the candidate matching images that best matches the combination of the plurality of captured images.